In the new round of Internet innovation, the large model is fortunate to become the next competition point, this time it is no longer as fleeting as the previous metaverse, but is really gradually penetrating into the application level.

According to the release of the 2024 China Mobile Internet Semi-Annual Report, the monthly active users of AIGC apps reached 61.7 million in June 2024, a year-on-year increase of 653%. Since last year, the large model has once again set off an Internet melee, and the global technology companies are afraid of missing an important turning point of the times, whether it is making games, e-commerce, or social players, they are doing their best to follow up the era of large models.

Some institutions have predicted that the global AI market will exceed $6 trillion by 2025, with a compound growth rate of 30% from 2017 to 2025.

It is true that the Internet world has been quiet for many years, and it has finally been lively again. Taobao, Alipay, Douyin and other top 20 super apps in domestic traffic are basically working on embedded AI applications, intelligent assistants, intelligent search, and intelligent shopping guides...... All kinds of new ways to play are emerging.

After the global Internet field loses its ability to innovate, can the big model really give new vitality to the Internet? This is a question worth pondering.

Innovation perception is downgrading, and product development is complacent

Before the emergence of large models, why have Internet giants not incubated new innovative products for a long time?

In fact, technology companies around the world are being brutally pressed to ask this question. Last January, the United Kingdom journal Nature published a paper based on 45 million manuscripts and 3.9 million patent discoveries that disruptive technologies are declining worldwide.

From the perspective of enterprises, over the years, the Internet market has never stopped its research and development progress, and even continued to strengthen. It's just that the R&D of large factories seems to have lost their meaning, and the long-term investment is not proportional to the recovery, which seriously consumes the innovation enthusiasm of capital.

This is not unfounded. According to the report of the Shanghai Stock Exchange, in 2021 alone, the R&D cost of a number of domestic science and technology innovation enterprises represented by Cambrian will be as high as 16.7 billion, but unfortunately, the cumulative loss far exceeds this number. According to data from the Bureau of Statistics, from 2000 to 2019, the investment of enterprises has exceeded 76% of the total domestic R&D investment, with an annual growth rate of 10%.

Even in the unfavorable external environment in recent years, the R&D growth rate of domestic enterprises can still be maintained at more than 18%. Previously, Baidu's Robin Li publicly stated that Baidu had more than 10,000 R&D engineers, and the investment was once 20% of the revenue, but the actual performance in exchange was not ideal.

In this way, giant companies prefer to invest in existing projects, taking Tencent as an example, data shows that Tencent currently has a total investment of more than 800 enterprises, of which 160 are unicorn companies with a valuation of more than $1 billion. For this reason, there has even been a saying that the investment and intervention of giants have compressed the innovation power within the industry.

In addition, R&D cannot be exchanged for considerable recycling, which also allows large manufacturers to no longer blindly develop new products.

In recent years, the frequency of the emergence of finished mini programs is much higher than that of independent apps, and previously, Alibaba, Tencent, Byte, Baidu, Kuaishou, Meituan, JD.com, ...... Mini programs have been developed one after another, and independent apps have been shut down on a large scale in order to save the cost of trial and error. Statistics show that Tencent has shut down more than 40 projects in a year, and Byte has also removed Party Island, so far, only the Internet manufacturers have shut down more than 70 independent products in recent years.

According to data from the Ministry of Industry and Information Technology, in the first quarter of this year, the operating costs of Internet enterprises above designated size in China increased by 5.1% year-on-year, achieving a total profit of 27.89 billion yuan, but a year-on-year decrease of 15.3%, and the growth rate of total profits turned from positive to negative.

The emergence of the large model is a ray of light shining into the Internet world, and the large manufacturers who are struggling to rest on their laurels are rushing up, and it can be seen from the direction of research and development that the large model has indeed inspired the research and development confidence of the giants. But how long can the innovation brought about by the big model last?

One thing to note is that today, it is difficult to have a phenomenal product or a leading technology in the Internet field where innovation is weak. After all, after the era of WeChat and Douyin, any bit of wind and grass can trigger the involution of the industry, just as at present, self-developed chips, big data, cloud computing, artificial intelligence and other technologies have become the highlight of all giants and even technology entrepreneurship.

The homogeneous drama has never disappeared in the Internet world, and when AI gameplay can be seen on any APP, such innovation is no longer "innovation".

In addition, although the craze for large models has caused some splashes, the problem of the disproportion between R&D and revenue that Internet giants once feared is even more serious. The global technology is driving AI, and more and more capital expenditures are being created. During this time, the financial reports of overseas giants vividly showed the essence of large models burning money.

According to institutional analysis, by 2025 and 2026, the cost of large model training will be close to 50-10 billion US dollars, of which Meta, Google, and Microsoft may plan to increase the cost of large model research and development to 50 billion US dollars.

There are signs that the Internet may never stop innovating, but the awareness of innovation has declined.

The power of the application of large models is not as strong as expected

Unlike several innovations in the past, this time the large model application provided by the Internet to users collectively encountered some troubles not long after it was launched: Is there a high probability that users will need a large model? Judging from the current set of data, the answer may be more pessimistic than expected.

According to data from Sequoia Capital, even ChatGPT, the head of the world's largest model, had a first-month user retention rate of only 56%, and about half of the users "shelved" it in less than a month. Similarly, the "2024 China Mobile Internet Semi-annual Report" also shows that domestic AIGC users are unstable, and the average time spent per capita in the AIGC industry has declined by 23.5% year-on-year.

In the final analysis, the penetration of artificial intelligence into real life is just a "fantasy" of capital.

From the user level, the user retention rate of AIGC apps on almost all mainstream apps is lower than that of traditional apps, and the engagement rate is also low. In July, Beike Finance released a survey, 52.05% of respondents sometimes use large models in their work, 23.97% rarely use them, 20.55% of respondents use them frequently, and only 2.05% of people use them all the time.

At the enterprise level, Huawei has a set of predictions that by 2026, the penetration rate of artificial intelligence in enterprises will only reach 20%.

Why is this happening? Technology, cost, usability, and security are all reasons.

Taking the entertainment industry, which is the most widely used AI landing, as an example, some time ago, Jackie Chan's new film "Legend" was released, and before the release of the film, AI technology has been the biggest gimmick in the promotion of the film. It is reported that Bona Film used AI to restore the 27-year-old Jackie Chan in the movie, but few viewers bought it.

The data shows that the current Douban score of "Legend" is 5.4 points, and the box office is only more than 70 million in more than ten days after its release.

Another major application area, "advertising", has also received mixed word-of-mouth. According to iResearch, about half of advertisers have applied AIGC technology in online marketing activities, of which more than 9% are used for content and creative scenarios.

However, AIGC's shortcomings have also begun to surface: for example, the production materials are too formulaic, the AI effect makes users aesthetically fatigued, and the well-known AI plagiarism problem...... Previously, the article "I used AI to generate an ad in five minutes, but it took five hours to taste AI" generated heated discussions on social platforms.

If the large model cannot produce a rigid demand effect in the user's online life like social communication and short video entertainment, then the large model will not have much of a role in promoting the Internet process. At present, the biggest focus of the Internet field is to improve the application efficiency of AI implementation.

Capital is also aware of this, and the flow of investment is flowing from the R&D track to the application track. According to Haitong International Research Report, 2024 is expected to be the first year of the full commercial implementation of domestic large models.

According to the data, among the nearly 120 global large-scale model investment events this year, large-scale model application companies accounted for 69%, accounting for more than half, while AI Infra and general large-scale models accounted for only 16% and 11% respectively, and only 3% of large-scale model data services remained. Looking at the application fields of large models, AI healthcare, vision/video generation, office assistants, and programming assistants are the most densely funded, accounting for 15%, 15%, 13%, and 11% respectively.

All in all, capital is accelerating the popularization of large models in the real world, and how to match technology and business needs is an urgent issue for large model enterprises. Only in this way can the Internet be "saved", and on the contrary, the Internet, which has lost its innovative power, will continue to be confused.

There is not much left on the Internet for large models to eat their old books

There is a problem that needs to be noted, the Internet has reached the stage of large models, and most of the gameplay is still the same as before, either continuing to fight price wars, or turning back to eat their own traffic "old capital".

In essence, the landing of the large model is no different from the "enclosure race" in the early Internet period.

In May this year, a number of domestic large model players began to officially announce price cuts, Ali's Tongyi Qianwen main large model Qwen-Long's API price dropped by 97%, Wenxin large model two main models ERNIE Speed and ERNIE Lite are all free, and then, iFLYTEK also announced that iFLYTEK Xinghuo API capabilities are free and open.

On the ByteDance side, it took only 30 days for Doubao to rush to the first place from its release, and it is reported that the reason why Doubao can become the "top stream" of the large model in a short period of time is not only because of the help of Douyin with 794 million monthly active people, but also the amount of money burned in the new round has reached 124 million yuan.

Back then, the most tried and tested trick of major domestic Internet companies was to throw money. To this day, does the "enclosure" style of play still apply?

First of all, the key to the large model in the current can only spend money for traffic lies in the convergence of technology, and the final impact on user retention will also be the return to technology, simply reducing the application cost can indeed increase exposure and compete for users from a short-term perspective, but in the long run, AI technology services are not takeaways, let alone short videos, and relying on burning money cannot bring a good user experience.

Secondly, the development of large models itself is a huge cost of capital engineering, perhaps for large factories with rich cash flow, the price war can be afforded, but the current large model profitability is far away, and the risk of small enterprises entering the market cannot be underestimated, which will inevitably further reduce the innovation and creativity of the entire industry.

In fact, the price war for large models started from overseas, when OpenAI and Google were the first to announce price cuts. But overseas, cloud vendors are moving away from traditional service models and turning to other ways to fill this cost, taking Nvidia as an example, in May, Nvidia released data for the first quarter of fiscal year 2025.

Nvidia said that training and inferring AI on Nvidia CUDA can drive growth in cloud rental revenue, and every $1 spent on Nvidia's AI infrastructure gives cloud service providers the opportunity to earn $5 in GPU instant hosting revenue over four years. Whether China can quickly follow up on this step of the plan is actually debatable.

Of course, in addition to the ability to continue to "inherit" the way of playing, the Internet has not left much "family foundation" for large models over the years, even from a global perspective, in addition to funds, there has been a shortage of information and data that large models need most.

Similarweb's data shows that ChatGPT's traffic growth has gradually slowed since it peaked at 1.8 billion global visits in May 2023. In response, OpenAI decided to relax restrictions on ChatGPT, and users could use it without registration at one point.

This is also one of the dilemmas of the current development of large models: the existing amount of Internet information is difficult to support the training of such large models.

During this time, the fact that Byte and a number of online office companies "feed" the large model has caused many users to be dissatisfied. According to public information, the amount of data involved in GPT-4 training is as high as 12 trillion tokens, and in the future, like GPT-5, it may need 60 trillion to 100 trillion tokens.

According to the Epoch Institute, there is a 50% chance that the demand for high-quality data from large models will exceed the supply by mid-2024, and a 90% chance that this will happen by 2026, and this risk of data shortage will be delayed until 2028.

As for how to make up for this huge data gap, the penetration rate of the Internet is gradually touching the ceiling, and there is no better way for a while.

This article is from Xinzhi self-media and does not represent the views and positions of Business Xinzhi.If there is any suspicion of infringement, please contact the administrator of the Business News Platform.Contact: system@shangyexinzhi.com